library_name	peft
license	other
base_model	LiquidAI/LFM2.5-1.2B-Instruct
tags	unsloth trl sft lora function-calling mobile-actions on-device lfm2 lfm2.5 hf_jobs
datasets	google/mobile-actions
language	en
pipeline_tag	text-generation

LFM2.5-1.2B-Instruct-mobile-actions

A LoRA fine-tune of LiquidAI/LFM2.5-1.2B-Instruct on google/mobile-actions for on-device function calling on Android.

The model translates natural language instructions into executable function calls for Android OS system tools (e.g. sending messages, setting alarms, making calls, web search).

Training Details

Parameter	Value
Base model	LiquidAI/LFM2.5-1.2B-Instruct (1.17B params)
Method	SFT with LoRA via Unsloth
Dataset	google/mobile-actions (8,693 train / 961 eval)
Hardware	NVIDIA L4 (22 GB VRAM) on HF Jobs
Training time	122.9 minutes
Epochs	3
Batch size	2 x 4 gradient accumulation = 8 effective
Learning rate	2e-4 (linear decay)
Optimizer	AdamW 8-bit
Max sequence length	2048
Precision	16-bit with Unsloth optimizations

LoRA Configuration

Parameter	Value
Rank (r)	16
Alpha	16
Target modules	q_proj, k_proj, v_proj, out_proj, in_proj, w1, w2, w3
Trainable parameters	11,108,352 / 1,181,448,960 (0.94%)
Dropout	0
Gradient checkpointing	Unsloth

Training Results

Metric	Value
Final train loss	0.0137
Epoch 1 eval loss	0.0161
Epoch 2 eval loss	0.0155
Epoch 3 eval loss	0.0154
Total steps	3,261
Speed	0.442 steps/s

Training was response-only (user/system tokens masked from loss). Eval loss tracks train loss closely with no overfitting.

Inference Results

Evaluated on 20 held-out examples from the google/mobile-actions eval split: 20/20 correct (100% accuracy)

#	User Prompt	Expected	Predicted	Time
1	Can you please save a new contact for me? The name is Lena Petrova...	create_contact	create_contact	30.41s*
2	Please send an email to javier.ortega@ecotradeintl.com with the subject 'Update...	send_email	send_email	5.31s
3	I need to save a new contact. The full name is Anya Sharma...	create_contact	create_contact	2.62s
4	Please set up a new calendar event for 'Team Lunch with Marketing' on May 13...	create_calendar_event	create_calendar_event	2.06s
5	Please send an email to Kenji Tanaka at kenji.tanaka@corpmail.jp...	send_email	send_email	3.15s
6	Turn on the flashlight and show me the location of the Sunnyvale Library...	turn_on_flashlight	turn_on_flashlight	1.82s
7	Can you please show me the location of the art supply store 'Canvas Creations'...	show_map	show_map	1.77s
8	I need to set up a new calendar event. The title should be "Meeting with Dr. Che...	create_calendar_event	create_calendar_event	2.16s
9	Please schedule a calendar event titled 'Quarterly Budget Review' for 10:30 AM...	create_calendar_event	create_calendar_event	4.64s
10	I need to check under the sofa, please turn on the flashlight.	turn_on_flashlight	turn_on_flashlight	0.90s
11	Please save a new contact. The name is Marcus Oliveira...	create_contact	create_contact	2.91s
12	Can you show me where The Wandering Page bookstore is located in Brooklyn?	show_map	show_map	1.24s
13	I'm having trouble reading this menu. Can you please turn on the flashlight...	turn_on_flashlight	turn_on_flashlight	1.86s
14	Send an email to Aisha.Khan@fabrikam.com with the subject "Follow-up on Q1 Repor...	send_email	send_email	3.03s
15	Please send an email to elara.pereira@email.com with the subject "Project Checkp...	send_email	send_email	4.93s
16	Please turn off the flashlight and immediately schedule a calendar event titled...	turn_off_flashlight	turn_off_flashlight	2.84s
17	Turn on my flashlight, and please create a calendar event titled "Morning Run"...	turn_on_flashlight	turn_on_flashlight	2.70s
18	Please create a new contact named Aisha Khan with the phone number 987-654-3210...	create_contact	create_contact	3.87s
19	I need to add a new contact named Elara Chen. Her phone number is +61 491 570 11...	create_contact	create_contact	2.45s
20	Turn on the flashlight, I need to see what I dropped in this dark corner.	turn_on_flashlight	turn_on_flashlight	0.89s

*First call includes CUDA warmup. Typical inference: ~1-5s on NVIDIA L4.

Accuracy by function type:

Function	Tested	Correct	Accuracy
create_contact	5	5	100%
send_email	4	4	100%
create_calendar_event	4	4	100%
turn_on_flashlight	4	4	100%
show_map	2	2	100%
turn_off_flashlight	1	1	100%
Total	20	20	100%

Usage

With PEFT + Transformers

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("LiquidAI/LFM2.5-1.2B-Instruct")
model = PeftModel.from_pretrained(base_model, "kshitijthakkar/LFM2.5-1.2B-Instruct-mobile-actions")
tokenizer = AutoTokenizer.from_pretrained("kshitijthakkar/LFM2.5-1.2B-Instruct-mobile-actions")

messages = [
    {"role": "system", "content": "You are a helpful assistant with access to the following functions. Use them if required.\n\n" + tools_json},
    {"role": "user", "content": "Send a text message to John saying I'll be late"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.1, top_k=50, top_p=0.1, repetition_penalty=1.05)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With Unsloth (faster inference)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="kshitijthakkar/LFM2.5-1.2B-Instruct-mobile-actions",
    max_seq_length=2048,
)
FastLanguageModel.for_inference(model)

Dataset Format

Each training example contains:

• System prompt: Available Android function definitions (JSON)
• User message: Natural language instruction (e.g. "Set an alarm for 7am tomorrow")
• Assistant response: Function call as JSON (e.g. [{"name": "set_alarm", "arguments": {"datetime": "..."}}])

The model was trained on responses only, learning to map user instructions to the correct function calls given available tool definitions.

Intended Use

• On-device function calling for Android system tools
• Translating natural language to structured API calls
• Edge deployment for mobile assistants

Limitations

• Trained only on Android system tool functions from the google/mobile-actions dataset
• May not generalize to arbitrary function schemas outside the training distribution
• Inherits limitations of the base LFM2.5-1.2B-Instruct model

Training Script

The training script is available at: github.com/Mandark-droid/LFM2.5-1.2B-Instruct-mobile-actions